Learning method, judgment method, reliability estimation device, and recognition device

Abstract

To estimate how feature quantities are suitable for identifying a class of input data, the feature quantities being obtained by inputting the input data to a feature quantity extraction unit.SOLUTION: A learning method of a reliability estimation apparatus that estimates a degree of reliability of feature quantities to be used for identifying a class of input data, the feature quantities being obtained by inputting the input data to a feature quantity extraction unit includes: inputting ground truth class data, which is data on a ground truth class, to the feature quantity extraction unit to extract feature quantities of the ground truth class data (S1); inputting the extracted feature quantities to a class identifying unit to calculate a ground truth reliability which is a value indicating a degree of certainty that the class identified by the class identifying unit is a ground truth class (S2); and training the reliability estimation apparatus so that a reliability to be output from the reliability estimation apparatus that has received input of the feature quantities of the ground truth class data may be coincident with the ground truth reliability (S3, S4).SELECTED DRAWING: Figure 4

Description

[Technical Field]

[0001] The present invention relates to a learning method, a determination method, a reliability estimation device, and a recognition device. [Background technology]

[0002] Patent Document 1 describes a recognition device that includes a feature extraction means for extracting features of an input image, a template feature storage means for storing template features of a specific class, and a classification means for identifying the class of the input image based on the features extracted by the feature extraction means and the template features. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent Publication No. 2021-117565 Summary of the Invention [Problem to be solved by the invention]

[0004] In order to achieve high recognition accuracy with the above-described recognition device, prior art has proposed inventions relating to feature extraction means. For example, in Patent Document 1, a teacher feature extraction model is prepared that is more accurate but requires more calculations than the feature extraction model of the recognition device, and learning is performed so that the features output by the feature extraction model match the features output by the teacher feature extraction model. In this way, a highly accurate feature extraction model with a small amount of calculations is obtained, and object recognition is realized. However, in order to achieve higher recognition accuracy, it is not enough to simply provide a highly accurate feature extraction means. It is desirable to make the template features stored in the template feature storage means the most appropriate template features for the class to be recognized.

[0005] However, the above-mentioned Patent Document 1 does not mention how to configure template features, and the generated template features are not necessarily template features generated in advance to be optimal for the class to be recognized (hereinafter, may be referred to as "optimal template features"). In this case, it is unclear to what extent the template features stored in, for example, the template feature storage means are suitable for class identification. The present invention has been made in consideration of the above-mentioned problems, and aims to estimate the degree to which features obtained by inputting input data to a feature extraction unit are suitable for identifying the class of the input data. [Means for solving the problem]

[0006] According to one aspect of the present invention, there is provided a training method for a reliability estimation device that estimates the reliability of features used to identify the class of input data obtained by inputting input data to a feature extraction unit. The training method includes inputting correct class data that is data of the correct class to the feature extraction unit to extract features of the correct class data, inputting the extracted features to a class identification unit to calculate a correctness reliability that is a value representing the likelihood that the class identified by the class identification unit is the correct class, and training the reliability estimation device so that the reliability output by the reliability estimation device to which the features of the correct class data have been input coincides with the correctness reliability. According to another aspect of the present invention, there is provided a reliability estimation device trained by the above-described training method. While the class identification unit calculates the accuracy reliability using the reference template feature, the reliability estimation device can output the reliability without using the reference template feature. Therefore, it can output the reliability for features other than the trained correct class.

[0007] According to yet another aspect of the present invention, there is provided a determination method for determining features extracted from features of input data. In the determination method, classification target class data, which is data on a class to be classified, is input to a feature extraction unit that extracts features of the input data, and features of the classification target class data are extracted, the reliability of the features extracted by the feature extraction unit is estimated by a reliability estimation device that has been trained by the above-mentioned learning method, and it is determined whether or not to adopt the features extracted by the feature extraction unit as template features for classifying the class to be classified, based on the reliability estimated by the reliability estimation device. According to yet another aspect of the present invention, there is provided a recognition device comprising a template feature employed by the above-described determination method and a feature extraction unit.

[0008] According to yet another aspect of the present invention, there is provided a recognition device including a feature extraction unit that extracts features of input data, and a class identification unit that identifies a class of the input data based on the features extracted by the feature extraction unit. The recognition device inputs the input data to the feature extraction unit to extract the features of the input data, estimates the reliability of the features of the input data using a reliability estimation device trained by the above-mentioned learning method, and changes the processing of the input data depending on the reliability. [Effects of the Invention]

[0009] According to the present invention, it is possible to estimate how suitable a feature obtained by inputting input data to a feature extraction unit is for classifying the input data. [Brief explanation of the drawings]

[0010] [Figure 1] 1 is a schematic diagram illustrating an example of the hardware configuration of a recognition device or a learning device according to an embodiment of the present invention. [Figure 2] FIG. 2 is a block diagram illustrating an example of a functional configuration of a learning device according to an embodiment of the present invention. [Figure 3] 1 is a block diagram illustrating an example of a functional configuration of a recognition device according to an embodiment of the present invention. [Figure 4]10 is a flowchart illustrating an example of a learning method for a reliability estimation unit according to an embodiment of the present invention. [Figure 5] 10 is a flowchart illustrating an example of a registration process for storing feature amounts of input data in a template feature amount storage unit in the recognition device according to an embodiment of the present invention. [Figure 6] 10 is a flowchart illustrating an example of a recognition process for identifying input data in the recognition device according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0011] Hereinafter, embodiments of the present invention will be described with reference to the drawings. Note that the embodiments of the present invention shown below are examples of devices and methods for embodying the technical idea of ​​the present invention 6, and the technical idea of ​​the present invention does not limit the structure, arrangement, etc. of the components to those described below. The technical idea of ​​the present invention can be modified in various ways within the technical scope defined by the claims. As an example of an embodiment of the present invention, an example of a recognition device will be described, which receives an image as input data, extracts features of the input image of a specific class (e.g., a person), and recognizes the class of the input image based on the extracted features. The recognition device may perform, for example, face recognition or person identification (Re-Identification: ReID). However, the present invention is not limited to this, and may also be applied to a recognition device that receives, for example, a voice signal as input data and recognizes the class of the input voice.

[0012] 1 is a schematic diagram showing an example of the hardware configuration of a recognition device 1 or a learning device 10 according to an embodiment of the present invention. The recognition device 1 and the learning device 10 may basically be configured with the same hardware configuration. The recognition device 1 and the learning device 10 may be realized by integrated hardware (the same hardware), or may be configured by separate hardware. The recognition device 1 and the learning device 10 include an imaging unit 2, a communication unit 3, a storage unit 4, an image processing unit 5, an output unit 6, and an operation input unit 7.

[0013] The image capturing unit 2 is a surveillance camera installed for the purpose of monitoring a predetermined area, and is attached in a position where it can capture images of people staying within the area. The images captured by the image capturing unit 2 are transmitted to the image processing unit 5 via the communication unit 3. The communication unit 3 transmits and receives data between the imaging unit 2, the image processing unit 5, the output unit 6, and the operation input unit 7. A LAN (Local Area Network) or a public line such as the Internet can be used.

[0014] The storage unit 4 is configured by a hard disk drive (HDD) or a solid state drive (SSD), and stores various programs including an operating system, and various data. The image processing unit 5 is composed of a CPU, MPU, peripheral circuits, terminals, various memories, etc., and transmits the results of image processing performed on the image captured by the imaging unit 2 to the output unit 6 via the communication unit 3.

[0015] The output unit 6 is a display, a projector, a printer, a removable drive, a USB (Universal Serial Bus) interface, a network interface, or the like that outputs various information generated by the recognition device 1. The operation input unit 7 is a mouse, keyboard, etc. that is operated by the user to accept inputs specifying templates and search ranges.

[0016] 2 is a block diagram of an example of the functional configuration of a learning device 10 according to an embodiment of the present invention. The learning device 10 is a device used for training a reliability estimation unit 16 that estimates an index value (hereinafter referred to as "reliability") that indicates how suitable the feature values ​​of input data are for classifying the class of the input data. In other words, the reliability is a value that indicates the likelihood that the class can be correctly identified when the class of input data is identified using the feature values. The learning device 10 includes a learning data acquisition unit 11, a feature extraction model storage unit 12, a feature extraction unit 13, a template feature storage unit 14, a class identification unit 15, a reliability estimation unit 16, a reliability estimation model storage unit 17, and a learning unit 18. The photographing unit 2 or the memory unit 4 in FIG. 1 functions as a learning data acquisition unit 11, the memory unit 4 functions as a feature extraction model memory unit 12, a template feature memory unit 14, and a reliability estimation model memory unit 17, and the image processing unit 5 functions as a feature extraction unit 13, a class identification unit 15, a reliability estimation unit 16, and a learning unit 18.

[0017] The learning data acquisition unit 11 acquires data of a known class (hereinafter sometimes referred to as a "correct class") as learning data. The number of correct classes may be one or more. For example, the learning data acquisition unit 11 may acquire data of a correct class that has been prepared in advance as learning data and stored in the storage unit 4 as learning data. The learning data acquisition unit 11 may also acquire, as learning data, an image obtained by capturing an image of the correct class with the imaging unit 2.

[0018] The feature extraction model storage unit 12 stores a feature extraction model. For example, the feature extraction model may be modeled as a CNN configured with a multi-layer network such as that used in deep learning. In this case, the feature extraction model has a network structure in which multiple layers, such as convolutional layers, activation functions, and pooling layers, are connected in series, and the feature extraction model storage unit 12 stores information including the filter coefficients of the filters that make up the network and the network structure. These filter coefficients and other parameters of each layer are updated through learning.

[0019] The feature extraction unit 13 extracts features from the training data using a feature extraction model. The feature extraction unit 13 reads the feature extraction model from the feature extraction model storage unit 12. The feature extraction unit 13 also receives training data from the training data acquisition unit 11. The feature extraction unit 13 inputs the training data to the feature extraction model and outputs the features output by the feature extraction model to the class identification unit 15 and the reliability estimation unit 16. Note that when the feature extraction model of the feature extraction unit 13 is configured with multiple layers of CNNs, the feature extraction unit 13 may output the features output from the CNN in the final layer to the reliability estimation unit 16, or may output the features output from any of the CNNs in the intermediate layers to the reliability estimation unit 16, or may output both the features output from the CNN in the final layer and the features output from the CNNs in the intermediate layers to the reliability estimation unit 16.

[0020] The class identification unit 15 reads out the optimal template features of each correct class from the template feature storage unit 14. The class identification unit 15 also receives the features extracted from the training data by the feature extraction unit 13. The class identification unit 15 compares the data received from the feature extraction unit 13 with the template features read from the template feature storage unit 14 to output a probability distribution indicating the probability that each class in the training data is the correct class. In the following description, the probability of a correct class output from the class identification unit 15 when a feature extracted from the training data of a certain correct class is received is referred to as the "true value reliability" of that feature. The true value reliability is an example of the "correct answer reliability" described in the claims. When the class identification unit 15 uses the optimal template features, features that can identify the correct class with a high probability have a high true value reliability, while features that can identify the correct class with a low probability have a low true value reliability. In other words, the true value reliability can be used as an index value that indicates how suitable a feature is for class identification. The class identification unit 15 outputs the true value reliability to the learning unit 18.

[0021] For example, the class identification unit 15 outputs a logit sequence a = (a1, a2, ...) (a = Wf) by multiplying the feature f by a matrix W = (w1, w2, ...) that represents the optimal template feature for each correct class Ck. Here, the logit ak represents ak = ln(p(x | Ck) p(Ck)). p(Ck) is the probability that the object to be identified belongs to the correct class Ck, and p(x | Ck) is the probability density distribution that the feature vector x is observed when measuring an object that belongs to the correct class Ck. Alternatively, the feature value f is multiplied by the matrix W to output a likelihood sequence p(x│.) = (p(x│C1), p(x│C2), ...) (p = Wf), and the logit sequence a is calculated from the logit definition formula above using a prepared prior distribution p(Ck). Then, we apply the softmac function p(Ck│x) = exp(ak) / (Σexp(aj)) to the logit sequence a to obtain the (posterior) probability sequence (i.e., probability distribution) p(.│x) = (p(C_1│x), p(C_2│x), …). The class identification unit 15 may output the logit or likelihood of the correct class to the learning unit 18 as the true value reliability instead of the probability of the correct class. For example, the class identification unit 15 may be modeled as a CNN configured with a multi-layer network such as that used in deep learning, so that the parameters of each layer, such as the filter coefficients of the network, can be changed by learning.

[0022] The template feature storage unit 14 stores the optimal template feature of the correct class. The optimal template feature may be received from an external device by, for example, the communication unit 3 and stored in the template feature storage unit 14. The optimal template feature may be calculated by learning. For example, features extracted from learning data of the correct class and the template feature of the correct class may be input to the class identification unit 15, and the template feature may be learned so as to increase the reliability of the output true value.

[0023] The reliability estimation unit 16 inputs the features of the learning data output from the feature extraction unit 13. When the feature extraction model of the feature extraction unit 13 is configured with multiple layers of CNNs, the reliability estimation unit 16 may input the features output from the CNN in the final layer, or may input the features output from any CNN in an intermediate layer, or may input both the features output from the CNN in the final layer and the features output from the CNN in an intermediate layer.

[0024] Furthermore, the reliability estimation unit 16 reads out a reliability estimation model stored in the reliability estimation model storage unit 17. The reliability estimation model is an estimation model that estimates the true value reliability output by the class identification unit 15 for the feature input from the feature extraction unit 13 without using a reference template feature. In the following description, the true value reliability estimated by the reliability estimation unit 16 will be referred to as "estimated reliability R." The reliability estimation unit 16 inputs the feature input from the feature extraction unit 13 into the reliability estimation model to estimate the estimation reliability R, and outputs the estimated estimation reliability R to the learning unit 18.

[0025] The reliability estimation model may be modeled, for example, as a CNN configured with a multi-layer network such as that used in deep learning. In this case, the reliability estimation model has a network structure in which multiple layers such as convolution layers, activation functions, and pooling layers are connected in series, and the reliability estimation model storage unit 17 stores information including the filter coefficients of the filters that configure the network and the network structure. These parameters of each layer, such as the filter coefficients, are updated through learning.

[0026] The learning unit 18 learns a reliability estimation model for the reliability estimation unit 16 so that the true reliability obtained by inputting the features of the correct class output by the feature extraction unit 13 to the class identification unit 15 and the reliability estimation unit 16 matches the estimated reliability R. Specifically, the reliability estimation model is learned so as to minimize the error of the estimated reliability R output by the reliability estimation unit 16 relative to the true reliability output by the class identification unit 15.

[0027] For example, the mean square error between the estimated reliability R and the true value reliability is calculated, and a parameter update amount for the reliability estimation model to reduce the error is calculated using a gradient method or a coordinate descent method with the error as an energy function, and the reliability estimation model is updated by the update amount, and the estimated reliability R is calculated again to evaluate the mean square error. This process is repeated until an iteration termination condition is satisfied. Here, the iteration termination condition may be, for example, that the error is equal to or smaller than a predetermined threshold value, or that the number of iterations (number of updates to the reliability estimation model) reaches a predetermined upper limit.

[0028] As a result, the reliability estimation unit 16 is trained to directly estimate true value reliability without using the optimal template feature. By using the reliability estimation unit 16 trained in this way, it is possible to output estimated reliability R for all features output from the feature extraction unit 13, regardless of whether the class of the input data input to the feature extraction unit 13 is included in the correct class of the training data.

[0029] Note that the learning unit 18 may learn the feature extraction model stored in the feature extraction model storage unit 12 and the template features stored in the template feature storage unit 14 simultaneously with learning the reliability estimation model. For example, the learning unit 18 may input features extracted by the feature extraction model from learning data of the correct class and the template features of the correct class to the class identification unit 15, learn the feature extraction model and the template features so that the output true value reliability is high, and simultaneously learn the reliability estimation model so that the estimated reliability R estimated by the reliability estimation model matches this true value reliability. The learning unit 18 may fix the feature extraction model (i.e., without learning the feature extraction model) and learn only the template features simultaneously with the reliability estimation model.

[0030] 3 is a block diagram of an example of the functional configuration of a recognition device 1 according to an embodiment of the present invention. The recognition device 1 is a device that acquires data of a class to be recognized (hereinafter referred to as "recognition target class") as input data, extracts features of the input data, and recognizes the class of the input data based on the extracted features and template features. The recognition device 1 includes an input data acquisition unit 21, a feature extraction model storage unit 22, a feature extraction unit 23, a template feature storage unit 24, a class identification unit 25, a reliability estimation unit 26, a reliability estimation model storage unit 27, a feature selection unit 29, and a classification result output unit 30. The photographing unit 2 in FIG. 1 functions as an input data acquisition unit 21, the memory unit 4 functions as a feature extraction model memory unit 22, a template feature memory unit 24, and a reliability estimation model memory unit 27, the image processing unit 5 functions as a feature extraction unit 23, a class identification unit 25, a reliability estimation unit 26, and a feature selection unit 29, and the output unit 6 functions as a classification result output unit 30.

[0031] The configurations of the feature extraction model storage unit 22, the feature extraction unit 23, the class identification unit 25, the reliability estimation unit 26, and the reliability estimation model storage unit 27 may be similar to the configurations of the feature extraction model storage unit 12, the feature extraction unit 13, the class identification unit 15, the reliability estimation unit 16, and the reliability estimation model storage unit 17 of the learning device 10 described with reference to Figure 2. For example, when the recognition device 1 and the learning device 10 are realized by integrated hardware (the same hardware), the feature extraction model storage unit 12, the feature extraction unit 13, the class identification unit 15, the reliability estimation unit 16, and the reliability estimation model storage unit 17 may be used as the feature extraction model storage unit 22, the feature extraction unit 23, the class identification unit 25, the reliability estimation unit 26, and the reliability estimation model storage unit 27.

[0032] The input data acquisition unit 21 acquires data of the recognition target class as input data. For example, the input data acquisition unit 21 may acquire an image obtained by capturing an image of the recognition target class using the capturing unit 2 as input data.

[0033] The reliability estimation model storage unit 27 stores the reliability estimation model learned by the learning device 10. When the recognition device 1 and the learning device 10 are realized by separate hardware, the learned reliability estimation model may be output from the reliability estimation model storage unit 17 of the learning device 10 by the output unit 6 and stored in the reliability estimation model storage unit 27 of the recognition device 1. Similarly, the feature extraction model may be output from feature extraction model storage unit 12 of learning device 10 via output unit 6 and stored in feature extraction model storage unit 22. The feature extraction model storage unit 22 stores the feature extraction model used in learning using learning device 10. The feature extraction model may be output from feature extraction model storage unit 12 of learning device 10 via output unit 6 and stored in feature extraction model storage unit 22.

[0034] The template feature storage unit 24 stores template features that the class identification unit 25 uses to recognize the class of the input data. For a class that is not included in the learning data (hereinafter, sometimes referred to as an "unlearned class"), optimal template features may not be provided. In this case, the template feature storage unit 24 of the recognition device 1 stores provisional template features of the unlearned class. When storing template features of an unlearned class in template feature storage unit 24, the user may operate operation input unit 7 in Fig. 1 to instruct execution of a registration process for storing the template features in template feature storage unit 24. When executing the registration process, data of the unlearned class is input to input data acquisition unit 21. Feature extraction unit 23 extracts features from the data of the unlearned class and stores the extracted features in template feature storage unit 24 as template features of the unlearned class.

[0035] As described above, the data of the unlearned class input to the feature extraction unit 23 may not be suitable for class identification. In this case, features suitable for the template features cannot be extracted. Therefore, the reliability estimation unit 26, which will be described later, may be used to calculate the estimated reliability R of the features and determine whether the features of the data of the unlearned class are suitable for the template features. For example, if the estimated reliability R is greater than a threshold value Th1, the features extracted from the data of the unlearned class are stored as template features in the template feature storage unit 24, but if the estimated reliability R is equal to or less than the threshold value Th1, they do not need to be stored.

[0036] Therefore, when the feature extraction unit 23 outputs a feature of input data of an unlearned class in the template feature registration process, the feature selection unit 29 selects the feature output by the feature extraction unit 23 in accordance with the estimation reliability R output from the reliability estimation unit 26 that inputs this feature. For example, when the estimation reliability R is greater than a predetermined threshold Th1, the feature output from the feature extraction unit 23 is stored as a template feature in the template feature storage unit 24. When the estimation reliability R is equal to or less than the predetermined threshold Th1, the feature output from the feature extraction unit 23 is not stored as a template feature in the template feature storage unit 24. This makes it possible to accurately determine features that are not suitable for class identification and determine whether or not to use them as template features, thereby reducing the rate of false recognition.

[0037] Next, when the recognition device 1 performs classification processing on input data, the data to be classified is input to the input data acquisition unit 21. The feature extraction unit 23 extracts features from the input data. If the estimation reliability R is greater than a predetermined threshold Th2, the feature output from the feature extraction unit 23 is input to the class classification unit 25. If the estimation reliability R is equal to or less than the predetermined threshold Th2, the feature output from the feature extraction unit 23 is not input to the class classification unit 25. As a result, the feature selection unit 29 determines whether or not to output a classification result for the class of this input data from the class classification unit 25, depending on the estimation reliability R of the features of the input data. This makes it possible to accurately determine feature amounts that are not suitable for class identification and determine whether or not to perform class identification, thereby reducing the rate of false recognition.

[0038] The classification result output unit 30 outputs the classification result of the class classification unit 25. That is, it outputs the recognition result of the class of the input data. When the feature selection unit 29 prohibits the input of features to the class classification unit 25, the classification result output unit 30 may output information that the class of the input data has not been recognized.

[0039] When the recognition device 1 and the learning device 10 are realized by integrated hardware (the same hardware), the feature extraction model of the feature extraction unit 23 and the class identification unit 25 may be trained simultaneously with the training of the reliability estimation model stored in the reliability estimation model storage unit 27. The feature extraction model may be fixed (i.e., the feature extraction model is not trained), and only the class identification unit 25 may be trained simultaneously with the reliability estimation model.

[0040] (Variation) In the above embodiment, an example has been described in which the feature selection unit 29 stops inputting features to the class identification unit 25 in accordance with the reliability estimated by the reliability estimation unit 26 (i.e., stops output of the class identification result of the input data by the class identification unit 25), but the present invention is not limited to this. The present invention can be applied to various other modified examples of the recognition device 1 as long as it changes the processing of the input data in accordance with the reliability estimated by the reliability estimation unit 26.

[0041] For example, the recognition device 1 may include a determination unit that determines whether or not to use the input data for class identification processing according to the reliability estimated by the reliability estimation unit 26. For example, the determination unit may determine that the input data will not be used for class identification when the reliability is lower than a predetermined value. For example, when it is determined that the input data will not be used for class identification, the feature extraction unit 23 may stop outputting features that the class identification unit 25 uses to recognize the class of the input data. Also, for example, when it is determined that the input data will not be used for class identification processing, the class identification unit 25 may stop outputting the identification result for identifying the class of the input data.

[0042] Furthermore, for example, if it is determined that the input data will not be used for class identification processing, the features extracted from the input data by the feature extraction unit 23 may be prohibited from being stored in the template feature storage unit 24 as template features.

[0043] (operation) FIG. 4 is a flowchart of an example of a learning method for the reliability estimation unit 26 according to an embodiment of the present invention. In step S1, the feature extraction unit 13 extracts the feature of the input data of the correct class included in the learning data. In step S2, the class identification unit 15 identifies the class of the input data based on the features extracted by the feature extraction unit 13 and the optimal template features of the correct class, and outputs a true confidence value indicating the probability that the class of the input data is the correct class. In step S3, the reliability estimation unit 16 receives the feature extracted by the feature extraction unit 13 and outputs an estimated reliability R, which is an estimate of the true value reliability output from the class identification unit 15 for this feature.

[0044] In step S4 , the learning unit 18 learns the reliability estimation model of the reliability estimation unit 16 so that the estimated reliability R output by the reliability estimation unit 16 coincides with the true value reliability output by the class identification unit 15 . In step S5, the learning unit 18 determines whether or not learning of the reliability estimation unit 16 is complete. If learning of the reliability estimation unit 16 is complete (step S5: Y), the process ends. If learning of the reliability estimation unit 16 is not complete (step S: N), the process returns to step S1.

[0045] FIG. 5 is a flowchart of an example of a registration process for storing the feature amounts of input data in the template feature amount storage unit 24 in the recognition device 1 according to the embodiment of the present invention. In step S10, the feature extraction unit 23 extracts the feature of the input data of the recognition target class acquired by the input data acquisition unit 21. In step S11, the reliability estimation unit 26 receives the feature extracted by the feature extraction unit 23 and outputs an estimated reliability R, which is an estimate of the true value reliability output from the class identification unit 25 for this feature.

[0046] In step S12, the feature selection unit 29 determines whether the estimation reliability R is greater than a predetermined threshold Th1. If the estimation reliability R is greater than the predetermined threshold Th1 (step S12: Y), the process proceeds to step S13. If the estimation reliability R is equal to or less than the predetermined threshold Th1 (step S12: N), the process ends. In step S13, the feature selection unit 29 stores the feature extracted in step S10 as a template feature in the template feature storage unit 24. Then, the process ends.

[0047] FIG. 6 is a flowchart of an example of a recognition process for identifying input data in the recognition device 1 according to an embodiment of the present invention. In step S20, the feature extraction unit 23 extracts the feature of the input data of the recognition target class acquired by the input data acquisition unit 21. In step S21, the reliability estimation unit 26 receives the feature extracted by the feature extraction unit 23 and outputs an estimated reliability R, which is an estimate of the true value reliability output from the class identification unit 25 for this feature. In step S22, the feature selection unit 29 determines whether the estimation reliability R is greater than a predetermined threshold Th2. If the estimation reliability R is greater than the predetermined threshold Th2 (step S22: Y), the process proceeds to step S23. If the estimation reliability R is equal to or less than the predetermined threshold Th2 (step S23: N), the process ends. In step S23, the feature selection unit 29 inputs the features extracted in step S10 to the class identification unit 25. The class identification unit 25 identifies the class of the input data based on the input features, and then the process ends.

[0048] (Effects of the embodiment) (1) The learning device 10 trains a reliability estimation unit 26 that estimates the reliability of features used to identify the class of input data obtained by inputting the input data to the feature extraction unit 23 of the recognition device 1. The reliability estimation unit 26 differs from the class identification unit 15, which calculates reliability by preparing template features for each correct class, in that it calculates reliability without using template features as reference. The feature extraction unit 13 of the learning device 10 receives correct class data, which is data of the correct class, and extracts features of the correct class data. The class identification unit 15 receives the features extracted by the feature extraction unit 13 and calculates a correctness reliability, which is a value representing the likelihood that the class identified by the class identification unit 15 is the correct class. The learning unit 18 trains the reliability estimation unit 16 so that the reliability output by the reliability estimation unit 16, to which the features of the correct class data are received, matches the correctness reliability. This makes it possible to estimate the reliability of features obtained by inputting input data to the feature extraction unit 23 of the recognition device 1, which indicates how suitable they are for classifying the input data. In particular, since the class identification unit 15 prepares a template for each correct class, it can output a high reliability for the correct class, but cannot output reliability for classes other than the correct class. In contrast, the reliability estimation unit 26 can output reliability without reference template features, and therefore can output reliability for classes other than the learned correct class.

[0049] (2) The reliability estimation unit 16 may be trained using at least one of the features extracted from the intermediate layer of the feature extraction unit 13 formed by a neural network of multiple layers and the features finally extracted by the feature extraction unit 13. This allows for more efficient selection of features for training the reliability estimation unit 16. Furthermore, the reliability estimation unit 16 can estimate the reliability of the features by also referring to the feature extraction process of the feature extraction unit 13. As a result, more detailed reliability estimation becomes possible.

[0050] (3) The feature extraction unit 13 and the class identification unit 15 may be trained so as to increase the accuracy reliability, and the reliability estimation unit 16 may be trained at the same time. Alternatively, the class identification unit 15 may be trained so as to increase the accuracy reliability, and the reliability estimation unit 16 may be trained at the same time. In this way, by training the feature extraction unit 13 and the class identification unit 15 simultaneously with the reliability estimation unit 16, it is possible to improve the learning efficiency.

[0051] (4) The recognition device 1 includes a reliability recognition unit 26 that has been trained by the learning device 10. It judges the features extracted from the features of the input data. The feature extraction unit 23 that extracts the features of the input data inputs the classification target class data, which is data on the classification target class, and extracts the features of the classification target class data. The reliability recognition unit 26 estimates the reliability of the features extracted by the feature extraction unit 23. Based on the reliability estimated by the reliability estimation device 26, the feature selection unit 29 judges whether or not to adopt the features extracted by the feature extraction unit 23 as template features for identifying the classification target class. This allows only features suitable for class identification to be selected and used as template features, thereby improving the recognition accuracy of unlearned classes even when optimal template features are not provided for unlearned classes not included in the training data.

[0052] (5) The recognition device 1 includes a feature extraction unit 23 that extracts features of input data, a class identification unit 25 that identifies the class of the input data based on the features extracted by the feature extraction unit 23, and a reliability estimation unit 26 trained by the learning device 10. The feature extraction unit 23 extracts features of the input data. The reliability estimation unit 26 estimates the reliability of the features of the learning target class data. The recognition device 1 changes the processing of the input data depending on the reliability estimated by the reliability estimation unit 26. This makes it possible to accurately determine whether or not to use features that are not suitable for class identification, thereby reducing the rate of false recognition. [Explanation of symbols]

[0053] 1...recognition device, 2...photographing unit, 3...communication unit, 4...storage unit, 5...image processing unit, 6...output unit, 7...operation input unit, 10...learning device, 11...learning data acquisition unit, 12, 22...feature extraction model storage unit, 13, 23...feature extraction unit, 14, 24...template feature storage unit, 15, 25...class identification unit, 16, 26...reliability estimation unit, 17, 27...reliability estimation model storage unit, 18...learning unit, 21...input data acquisition unit, 29...feature selection unit, 30...identification result output unit

Claims

1. A learning method for a reliability estimation device, which estimates reliability of a feature obtained by inputting input data into a feature extraction unit, the feature being used to identify a class of the input data by a class identification unit, and outputs the estimated reliability, The processor: inputting correct class data, which is data of the correct class, to the feature extraction unit to extract the feature of the correct class data; inputting the extracted feature amount of the correct class data into a class identification unit provided separately from the reliability estimation device; training the reliability estimation device so that a correctness reliability calculated by the class identification unit as a value representing the likelihood that the class identified by the class identification unit is the correct class coincides with the reliability output by the reliability estimation device to which the feature quantity of the correct class data has been input; A learning method characterized by:

2. 2. The training method according to claim 1, wherein the processor trains the reliability estimation device using at least one of features extracted from an intermediate layer of the feature extraction unit formed by a neural network of multiple layers and features finally extracted by the feature extraction unit.

3. 3. The learning method according to claim 1, wherein the processor trains the feature extraction unit and the class identification unit so as to increase the accuracy reliability, and at the same time trains the reliability estimation device.

4. 3. The learning method according to claim 1, wherein the processor trains the class identification unit and the reliability estimation device simultaneously so as to increase the accuracy reliability.

5. A reliability estimation device trained by the training method according to any one of claims 1 to 4.

6. A method for determining a feature extracted from input data, comprising: The processor: inputting classification target class data, which is data of a classification target class, into a feature extraction unit that extracts feature amounts of input data, and extracting feature amounts of the classification target class data; The reliability estimation device trained by the training method according to any one of claims 1 to 4 estimates the reliability of the feature extracted by the feature extraction unit; When the reliability estimated by the reliability estimation device is greater than a threshold, it is determined that the feature extracted by the feature extraction unit is to be adopted as a template feature for identifying the class of the object to be identified, and when the reliability estimated by the reliability estimation device is equal to or less than the threshold, it is determined that the feature extracted by the feature extraction unit is not to be adopted as the template feature. A determination method characterized by:

7. A recognition device comprising: the template feature employed by the determination method according to claim 6; and the feature extraction unit.

8. A recognition device comprising: a feature extraction unit that extracts features of input data; and a class identification unit that identifies a class of the input data based on the features extracted by the feature extraction unit, inputting input data to the feature extraction unit to extract features of the input data; The reliability of the feature quantities of the input data is estimated by the reliability estimation device trained by the training method according to any one of claims 1 to 4, When the reliability is lower than a threshold, the class identification unit stops identifying the class of the input data. A recognition device characterized by:

Images